Open datasets · Hugging Face Hub
Open research datasets.
Open research datasets for low-resource NLP, retrieval evaluation, safety, and multimodal reasoning. Each collection documents its source, annotation, and verification method.
284
Source publications
Govt. agricultural & land records
6
Regional dialects
Content tokens frozen
CC-BY-4.0
Open licensing
Permissive academic use
Per-dataset
Verification method
Documented below each entry
Open-Access Dataset Repositories (5 Collections)
Hugging Face HubKrishokChat — Bengali Agricultural Safety & Chemical Advisory
The full 85,979-instance benchmark: General QA (28,993), Treatment QA with chemical traces (11,224), Safety Refusal & Re-query (20,112), Table QA (25,650), plus a 1,000-query real-world farmer benchmark — from 284 official publications in 6 dialects.
AgriTrust-RAG — Retrieval Nodes, Entities & Queries
2,882 provenance-annotated knowledge nodes, 19,768 entities, 17,501 factual triples, and 1,000 annotated queries (300 farmer-anchored, 400 KG-grounded, 200 safety-critical) from 284 official agricultural publications (1999-2024). Companion to arXiv:2608.14886.
TRACE-BN — Structured Bangla-English Tutoring Traces
4,099 curriculum-guided 7-field tutoring traces (word glosses, literal/natural translation, contrastive grammar, learner error, practice Q&A) for CEFR A1-A2 learners from NCTB Classes 9-10. Teacher: Gemini 3.5 Flash Lite; audited by 3 educators (4.71/5). Companion to arXiv:2608.15223.
ChitraMiti-12.8k + NCTB-500 — Bengali Geometric Reasoning
12,874 synthetic Bengali planar geometry problems with 15-attribute structured descriptions, plus 500 real textbook diagrams (NCTB Classes 6-10). Three-phase protocol with adversarial consistency checks. Companion to arXiv:2609.12509.
KhatianDoc — Bengali Legal Land Records (Redacted)
1,634 QA pairs over 107 real RS Khatian records (Vumi Office, Munshiganj) with Ana-Ganda base-16 fractions. 100% human-transcribed, verified line-by-line by a land-law practitioner, anonymized with positional tokens, redacted images released. Companion to arXiv:2609.03597 (NLLP @ EMNLP 2026).
How each dataset was built
Verification method differs by collection, so it is documented per collection:
| Collection | Source | Annotation & verification | Size | Paper |
|---|---|---|---|---|
| KrishokChat | 284 government publications, 13 institutions | Answers extracted verbatim from cited sources; chemical traces; validation pipeline discards, never corrects | 85,979 instances + 1,000 farmer queries | arXiv:2606.29243 |
| AgriTrust-RAG | 284 agricultural publications, 1999-2024 | Provenance injected deterministically; 3-annotator audit (κ 0.81); closed-loop refinement | 2,882 nodes · 19,768 entities · 1,000 queries | arXiv:2608.14886 |
| TRACE-BN | NCTB English curriculum (Classes 9-10) | Teacher-model generation, then 3 automated filters + 3-educator audit (mean 4.71/5) | 4,099 traces | arXiv:2608.15223 |
| KhatianDoc | 107 real RS Khatian records, Vumi Office | Zero OCR: human transcription + independent land-law practitioner review; positional-token anonymization | 1,634 QA pairs (300 eval) | Preprint |
| ChitraMiti | DeepVision-103k filter + NCTB textbooks | 3 bilingual annotators audit (κ 0.81 / 0.78); 3.0% correction rate | 12,874 + 500 items | Preprint |
Data ethics & privacy
These datasets follow academic data-ethics practice. Where real-world documents contain sensitive personal data:
Positional Token Masking
In legal deed benchmarks (KhatianDoc), real citizen names are replaced with positional identifiers ([PERSON_1], [PERSON_2]) to protect citizen privacy while preserving multi-hop inheritance co-reference.
Verbatim Reference Extraction
In chemical safety benchmarks (KrishokChat), ground-truth answers are extracted verbatim from official government extension manuals rather than synthetic LLM generation, preventing synthetic error propagation.
from datasets import load_dataset
# 1. Load KrishokChat Agricultural Safety & Chemical Benchmark
dataset = load_dataset("RaiyanKhaan/krishokChat")
print("Tracks available:", dataset.keys())
print("Test sample:", dataset["test"][0])
# 2. Inspect structured chemical traces for dosage audit
sample = dataset["test"][0]
if "chemical_trace" in sample:
print("Dosage specification:", sample["chemical_trace"])Read the associated papers
Each dataset is paired with an empirical paper analyzing failure modes, baseline sweeps, and ablations.